Wednesday 11th October 2017
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Machine Learning and the Art of Causal Assumptions
Causal inference requires assumptions, some of which will be untestable. Domain knowledge is essential, at the very least to validate whether our data contains the appropriate measurements of interest. However, no domain knowledge will ever be perfect, and it is of interest to derive the consequences of assumptions expressed at different levels of specificity. This provides a whole continuum of trade-offs between domain-free and domain-specific assumptions that should be negotiated. This talk will discuss a toolbox of machine learning methods to address important structural assumptions used in practical causal models: to what "degree" some of our variables act as a surrogate experiment? To what "degree" are the common causes between a treatment and an outcome accounted for? We will see that these are neither black-and-white concepts, nor there is an easy and fully tractable way of answering these questions by domain-specific assumptions. Machine learning provides a modicum of automated reasoning that allows us to better explore competing causal models in a way that generalizes common tools for sensitivity analysis.